# Hugging Face releases Phonon TTS: Apache-2.0 on-device speech model with 10+ inference backends

> **Key takeaway:** Phonon is a 31-40 million parameter on-device text-to-speech model with Apache 2.0 commercial licensing, ranking first on Hugging Face's Open TTS Leaderboard for German and Spanish intelligibility while supporting CPU, GPU, and NPU inference across more than 10 backends.

**Published:** 2026-11-05T12:02:42.416915+00:00 
**Source:** Hugging Face Blog & Announcements 
**Category:** ai-open-models 
**Canonical URL:** https://fosswire.org/news/hugging-face-releases-phonon-tts-apache-20-on-device-speech-.html 

## Summary
Hugging Face has officially launched Phonon TTS, a compact text-to-speech model designed to run entirely on local devices without network connectivity. The release includes two model sizes at approximately 65 MB and 75 MB, supporting English, French, German, Spanish, and Portuguese with over 100 built-in voices and custom voice cloning capabilities. Phonon 31m, which speaks English, is released under the Apache 2.0 license permitting commercial use, while Phonon 40m covering all five languages uses a CC BY-NC non-commercial license with user-based commercial pricing.

The model is built on the Pocket TTS lineage and employs flow matching for audio generation rather than conventional TTS modeling approaches. Phonon achieved top rankings on Hugging Face's Open TTS Leaderboard, placing first in German and Spanish intelligibility and in the top three across all evaluated languages despite being up to 100 times smaller than competing models. It also leads CPU streaming benchmarks for time-to-first audio latency. The dedicated ptts inference framework supports more than 10 hardware backends including Intel and ARM CPUs, CUDA, Vulkan, Metal, WebGPU, Apple Core ML, and Snapdragon NPUs.

## Architecture, inference framework, and Transformers integration
- **Flow matching audio generation:** Phonon uses flow matching rather than traditional TTS modeling techniques, which enabled high-quality output from a compact architecture but introduced numerical instability during training that required specialized techniques to resolve across all five supported languages.
- **ptts inference framework:** A purpose-built inference engine supporting 10+ backends including CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU), and NPU (Apple Core ML, Snapdragon), optimized for single-stream on-device performance and minimal time-to-first audio.
- **Autoregressive latent generation:** The model generates audio latents autoregressively where each step's output becomes the next input, with training loss computed directly on latents; the codec mapping audio to latents functions as an audio tokenizer analogous to text tokenization in language models.
- **Transformers integration from v5.21:** Available as `GradiumPhononTTSForConditionalGeneration` in Hugging Face Transformers, serving as the reference PyTorch implementation with support for batched generation and full-graph `torch.compile` for GPU throughput at scale, with Pocket TTS added as the base and Phonon implemented as a 200-line modular delta.
- **Model parameter counts:** Phonon 31m contains 30,788,354 parameters with a separate audio tokenizer of 10,305,313 parameters, placing total active weights at approximately 41 million for the English model.

## Why it matters
Phonon provides a fully open-weight, commercially licensable on-device speech synthesis option that reduces dependency on cloud APIs for voice applications in Linux, mobile Linux, and embedded systems. The broad backend support including Vulkan and CPU inference makes it deployable across diverse open-source hardware from Raspberry Pi to RISC-V boards without vendor-specific acceleration requirements. For developers building privacy-respecting applications, offline-capable assistive technology, or localized game audio pipelines on Steam Deck and similar Linux handhelds, Phonon offers a benchmark-leading alternative to proprietary cloud TTS services.